High-Level Features for Image Indexing and Retrieval

Text/Ontology

The fist technique is text-based meaning that the image content is described in terms of textual keywords. This description can be manually provided to the system or obtained using a vocabulary, an “object-ontology” which provides a qualitative definition of high-level concepts from the low-level features. For example, an image region can be assigned the keyword “sky” if it is an “upper, uniform and blue-collared region.” The “blue” attribute can be itself obtained from low-level features concerning color distribution (e.g. an average RGB colour components between predefined thresholds that characterize a generic sky region). Examples of this kind of descriptors, not limited to colour but applied also to other low-level features, can be found in Ravishankar et al. (1993). Since an ontology is naturally hierarchical, that is, it not only contains a set of concepts but also relations between these concept, an object can receive multiple keywords based on its nature. Moreover, the keywords vocabulary can refers non only to objects within the image but also to the image itself as a whole. In this case, the attributes or concepts of interest are global ones (e.g. the image category such as “sunset,” “landscape,” “indoor,” “close-up,” etc.). An example of how an image can be hierarchically and globally described is ImageNet (Deng et al. 2009), an image database organized according to the WordNet (Miller et al. 1990) hierarchy, in which each node of the hierarchy contains hundreds and thousands of related images. Using a vocabulary of keywords, a query to retrieve similar images, can be performed by using traditional, text-based, information retrieval techniques.